Kimi K3's vision encoder, made language-searchable. A small trained projector maps Kimi K3's frozen MoonViT-V2 vision features into the fusion-embedding-2 text space, so K3's own visual features become searchable in plain language and sit in the same 2048-d space as fusion-embedding's text, image, video, and audio.
Only the ~401M vision tower is pulled from moonshotai/Kimi-K3 on first use, never the 2.8T
language model. This worker embeds images; query the vectors with text using the
fusion-embedding-2 endpoint.
One-click from the RunPod Hub.
Kimi-K3 is a gated model. Set HF_TOKEN on the endpoint to a Hugging Face token that has
accepted the Kimi K3 license, otherwise the MoonViT-V2 tower download fails.
curl -s https://api.runpod.ai/v2/<ENDPOINT_ID>/runsync \
-H "Authorization: Bearer $RUNPOD_API_KEY" \
-H "Content-Type: application/json" \
-d '{"input": {"image": "https://example.com/photo.jpg"}}'Returns a 2048-d embedding in the fusion-embedding-2 text space. To search, embed a text query with fusion-embedding-2 and take the cosine similarity.
This pack is one of the modalities Engram searches. Engram is the open cross-modal memory layer for physical AI: it indexes a robot's video, audio, and motion into one embedding space and answers questions about it in plain language, including temporal reasoning that retrieval alone cannot.
pip install engram-robomemRepo: https://github.com/Eximius-Labs/engram | PyPI: https://pypi.org/project/engram-robomem | Playground: https://www.eximiuslabs.com/playground
The projector is released under CC-BY-NC-4.0. The Kimi-K3 base is under the Kimi K3 License (permissive, notice required).